Hardening checklist
Preview. A checklist for operators, drawn from the security model; read that page for the reasons. Items marked (documented) are stated in the deployment READMEs; the rest follow from the code. None of this makes mutually distrusting users safe on one deployment: the per-session isolation design (MI1) is not built, so give each trust domain its own deployment.
On every path​
- Set
SH_REQUIRE_AUTH=true. Compose defaults tofalseand the operator'sANTHROPIC_*variables then serve unauthenticated turns for anyone who can reach port 8080. The VM template and Kubernetes manifest already set it. If you must leave it off for a demo, keep the port off any shared network. - Decide the tenancy.
MOCA_TENANCY=multimakes the worker refuse local tools, scrub ambient provider credentials and require auth. No shipped deployment sets it. The VM README says not to rely on it to separate two users (documented). - Keep
SH_ALLOW_OPERATOR_FALLBACKoff unless a one-key trial needs it. The control plane has no user allowlist, so with the fallback on any GitHub user who can reach it can spend the operator's key. - Keep
SH_ADMIN_SUBJECTSminimal. Empty means no admins. Admin can list other users' sessions and delete any bundle. - Back up the credential KEK ring before anything else. Losing it makes every stored credential unreadable (documented). Setup never rotates it. See Redis backup and recovery.
- Do not publish Redis, the supervisor admin port (8081), the relay exec listener or the control plane port to networks you do not trust. The admin port is unauthenticated.
- Keep the supervisor reachable at layer 4 only. No Ingress, edge Route or mesh sidecar in front of it.
- Pin images by digest. The harness and remote-worker images are
:latestby default; pass--image <ref>@sha256:...(harness) and--sandbox-image <ref>@sha256:...(remote worker) tosetup.sh, orSANDBOX_IMAGEon the VM. Redis is digest-pinned only on Kubernetes. Images are not signed. - Treat the supervisor-to-control-plane hop as plaintext. It carries the exchange bearer and session tokens, and in direct mode the real model key. Keep it on a network you control.
- Rotate tokens deliberately. Access tokens last 15 minutes and are never revoked, so shorten
SH_API_TOKEN_TTL_SECONDSif you need faster revocation.
Kubernetes (deploy/k8s)​
- Use a CNI that enforces NetworkPolicy (kind 0.24 or later, or OVN-Kubernetes). On one that does not, every isolation rule is open (documented). Run the smoke test's isolation claim to check.
- Look at your node network. If it uses publicly routable addresses, the sandbox egress rule does not block
it; set
SH_SANDBOX_EGRESS_EXCEPT(documented, #446). - Scope the admin port. Any pod in
mocacan read the unauthenticated/metrics. On the single-namespace OpenShift target (ocp-single) that includes the sandbox pods. Do not run unrelated workloads in that namespace, or tighten the policy. - Restrict who can read secrets in
moca. All four MU1 values (signing key, public keys, KEK, exchange token) are keys of one Secret,moca-mu1. - Delete a leftover
moca-genkeyspod ifsetup.shwas killed; its log holds the generated keys. - Consider encryption at rest for Secrets. This repository does not configure it, and per-user credentials sit in Secrets (sealed under the KEK, with metadata in plaintext annotations).
- Add what the manifests omit: no ResourceQuota, LimitRange, PID or ephemeral-storage limits, and no PodDisruptionBudget.
- On OpenShift, supply a real certificate (
--tls-cert/--tls-key, or--tls-secreton theocp-singletarget only) instead of the self-signed fallback, and remember the control plane's Route terminates TLS at the router.
Single VM (deploy/vm)​
- Restrict access to ports 8080 and 8090. Use the README's SSH
Match Userblock, or a firewall allowlist. An allowlist sends tokens in clear. The supervisor binds all interfaces regardless of the sandbox firewall, as does the relay attach port (9443); the sandbox firewall restricts only the sandbox bridge. - Close cloud metadata at the platform (for example IMDSv2 with a hop limit of 1). Container sandboxes can currently reach it (#357).
- Verify
isolate=strictis enforced. Setup refuses a podman that rejects the option but cannot tell one that stores it without enforcing it. It was verified on netavark 1.17.2. - Remember the shared user. The three units run as one account, so a supervisor compromise reads the control plane's credentials (documented).
- Redis is on loopback with no password and no named volume.
- If you add the microVM tier, run its worker under a hardened unit; the shipped one runs as root with no systemd sandboxing. Keep its relay token separate (the installer does this).
Docker Compose (deploy/compose)​
- Treat it as a trial. Redis has no password, sandboxes have internet access, and there are no resource limits.
- Set
SH_REQUIRE_AUTH=trueas soon as you add the control plane, and removeANTHROPIC_*from.envif you want no anonymous spend. - Publish ports on loopback only (the shipped file already does for the supervisor and control plane).
MicroVM tier (deploy/microvm)​
- Use it where you need network isolation, not user isolation. The guest has no network device and reaches only
moca-egress's allowlist, but nothing binds a workspace key to a user. - Guard the relay exec token. Whoever holds it can target any workspace.
- Never give the microVM worker the exec token. It uses its own attach token.
- Plan for silent workspace resets (#338) and for
SH_MAX_COMMITTED_MB, which gates memory but not workspace disk.
Reporting a problem​
Report vulnerabilities privately; see SECURITY.md.